Articles
Why NOVA?
The case for decentralized AI drug discovery
Our north-star thesis is that the long-term value of decentralized intelligence will come from tackling hard problems with solutions that impact humanity at a global scale.
We believe drug discovery is the highest-leverage opportunity to prove that thesis. And in the case of Bittensor specifically, we believe the subnets that endure will necessarily be the ones producing solutions the world outside the network is willing to pay for. Here is why we think NOVA is one of them.
NOVA 2026 H1: From the Network to the Lab Bench
Year 1 proved decentralized AI drug discovery works. In the first half of 2026 we focused on closing the loop — turning nonstop competition into better data, durable incentives, and now physical compounds tested in the lab.
Here's where Metanova Labs stands at the midpoint of 2026.
Improving Hit and Hit-to-Lead identification with NOVA Compound’s new scoring function
Heavy Normalization strengthens Boltz-2 predictions without changing the model itself.
Key outcomes:
AUC ≈ 0.81 for hit identification
~50% improvement in high-affinity recovery
~75% reduction in low-affinity contamination
Best performance on 4 of 6 diverse targets
The method adds no meaningful computational overhead beyond heavy-atom counting, yet delivers material improvements in screening efficiency and downstream decision-making.
Closing the Loop: From virtual to real world discovery
In less than a year, NOVA scaled from a proof-of-concept into a live incentive engine powering both molecular discovery and search algorithm optimization. The system runs continuously, coordinating global contributors and adapting in real time.
NOVA is designed to plug directly into real discovery workflows. It generates assets (molecule libraries, benchmarks, and early-stage leads) that can be licensed, validated, co-developed, or accessed via Discovery-As-A-Service. This represents a fundamentally different model in an industry still dominated by single, fragile bets.
TREAT-2: Fine-Tuned Epigenetic Target Prediction
TREAT-2: fine-tuned from PSICHIC, our base model, using a highly curated dataset of histone deacetylase (HDAC) interactions.
Incorporates both human and rat data to enhance cross-species translation—vital for preclinical pipeline progression.
Achieved a 23% drop in Mean Average Prediction Error and a 45% reduction in variance across validation folds signaling a leap in the model’s ability to generalize across novel chemical scaffolds. This means TREAT-2 doesn’t just memorize known interactions, it can predict new ones, dramatically increasing our chances of finding real therapeutic hits.
TREAT-1: Advanced Binding Predictions for DAT, SERT, AND NET
Protein–ligand affinity prediction is crucial for accelerating drug discovery, but most leading models depend on scarce and expensive 3D structural data.
PSICHIC: A state-of-the-art, sequence-only graph-attention network that incorporates physicochemical constraints, offering a structure-agnostic alternative is highly versatile. But, PSICHIC’s performance on specialized transporter targets like dopamine (DAT), serotonin (SERT), and norepinephrine (NET) required further refinement.
TREAT-1 demonstrates +79% enrichment factor compared to the baseline PSICHIC model, paving the way for more accurate and efficient drug discovery targeting critical neurotransmitter systems.
whitepapers
The following whitepapers chart the development of NOVA powered by Bittensor Subnet 68. From early architectural decisions to advanced incentive mechanisms, each release reflects a distinct stage in our push to build a self-improving, adversarial, and model-agnostic drug discovery engine.
While these versions document important conceptual shifts, the protocol continues to evolve rapidly. For the most current implementation and community coordination, check our GitHub and Discord channel.
tracking the evolution
V1 - March 2025
Focus: Core architecture, deterministic scoring, and single-molecule optimization
This foundational version introduces NOVA as a decentralized competition layer for virtual screening. Miners submit candidate molecules from the SAVI 2020 library (~1.75B compounds) and are rewarded based on binding affinity scores predicted by PSICHIC, a GNN-based oracle. The system runs winner-takes-all challenges on fixed protein targets, establishing the core validator-miner feedback loop and emphasizing adaptive search over brute force enumeration.
V2 - april 2025
Focus: Multi-molecule submissions, adversarial incentives, and chemical diversity metrics
This whitepaper outlines the Shannon Upgrade, which introduces a multi-molecule challenge format, randomized target–antitarget assignments, and frequency-adjusted scoring. Submissions are now evaluated not only for potency but for their internal structural diversity, calculated using MACCS key fingerprints and a full Shannon diversity index. This model-agnostic upgrade increases pressure on miners to innovate, stress-tests prediction models, and enhances the breadth and quality of molecular discovery.
V3 - May 2025
Focus: Epoch-based competition, anti-target penalties, and entropy-weighted scoring
This update introduces weekly target locks, hour-scale epochs, and a refined scoring framework. Submissions now include 100 molecules per epoch, and a hard duplicate-invalidation rule ensures novelty across the target-week. Scoring incorporates target vs. anti-target differentials and a decaying Shannon entropy bonus, which rewards early-stage chemical diversity. The system now better reflects real-world lead optimization dynamics and discourages redundant exploration.
